← All lessons
027

Synthetic Data

When real data runs out or is too expensive to label, models generate their own training data. Done naively — each generation trained purely on the previous one's output — it fails in a characteristic way called model collapse: common patterns get amplified, rare cases vanish, small errors compound, and after a few generations the distribution shrinks to a drifting blob. Done carefully it works and powers frontier training: filter the synthetic samples for quality, keep real data in the mix, and use a stronger model or a verifier as the source of truth.

Step through generations. In naive mode the cloud shrinks and drifts away from the real distribution (dashed) — edges die first. Turn on curation + real-data mixing and watch the same generations hold their spread.

Generationreal data
dashed = spread of the original real datadiversity remaining: 100%

Generation 0: the real distribution, in all its variety — including the rare cases at the edges.

Check yourself

Why do the rare edge cases disappear first when a model trains on its own output?

Go deeper (free): Nature (open access) — AI models collapse on recursive data

Next: Curriculum & Data Mixes